Inspectable AI & learning systemsPublic field guide

Yodi: a guide to inspectable AI and governed memory

Yodi explores small AI systems whose learning, memory, inputs, and outputs can be inspected. Its work includes specialized models, explicit evaluations, and proposals that another application can review before accepting. The aim is to make intelligence something a team can study and shape, with each claim tied to a particular experiment.

Demon Hunter Labs4 min readReviewed
Start with a practical evaluation

The useful starting point

What Yodi gives your team

Your team can examine how a result was learned, retrieved, or proposed, then test the specific mechanism against its own requirements.

Who this is for

AI engineers, applied research groups, and company or government R&D teams investigating small models and inspectable learning.

  1. 01

    Define a task

    Choose an input, the expected behavior, and a frozen evaluation.

  2. 02

    Inspect the mechanism

    Separate learned model behavior from retrieval and deterministic rules.

  3. 03

    Compare candidates

    Measure new abilities, retained behavior, and failures before selection.

Separate intelligence from supporting machinery

An answer can come from a trained model, a retrieved memory, or a rule. Those mechanisms have different strengths and failures. Yodi's design makes their roles explicit through data provenance, model and checkpoint records, documented neural controls, and attribution.

This distinction matters when evaluating a compact system. A successful lookup does not by itself prove that a model learned a concept. A model answering its training examples does not establish performance on unfamiliar language. A useful trial reports each route separately.

Learning should preserve what already worked

New learning is only part of the decision. A candidate may gain an answer while forgetting an old one, fail when a question is reworded, or improve a development set without passing a held-out check. Yodi's current work treats candidate preview, retained behavior, and selection as separate steps.

For a company, this creates a useful evaluation pattern: freeze a baseline, introduce a bounded change, and compare both new and previous tasks. Product integration adds another boundary. A typed proposal still needs the receiving application to decide whether and how to apply it.

Current public evidence

Verified bounded intelligence

Executable proofs cover rights-bearing data, float training, compact model and tokenizer tournaments, exact checkpoint resume, a 6,000-byte sequence model, 512-byte plastic reasoning, owner-filtered DynaCapDB memory, six modality families, ordered serving/backpressure, real Vulkan kernels, native Kyber, exact backend agreement, and a manifest that rejects undocumented neural levers. General intelligence and OS-wide control remain North Stars.

Explore the project and its evidence

Try one bounded question

Teach a small task without hiding the regressions

A suggested evaluation for your team.

  1. Select a documented model route and record its checkpoint and data provenance.
  2. Prepare a small task with reference answers and unseen rewordings.
  3. Record baseline behavior before any training or memory change.
  4. Evaluate a separate candidate on new questions and previously solved ones.
  5. Restart the candidate, repeat the probe, and inspect which outputs came from memory.
What a useful result looks like

A comparison that names the learned change, the retained abilities, and every relevant failure. Promote a candidate only when the defined task and retention requirements pass.

Before you go further

Common questions about Yodi

Is Yodi a general replacement for a frontier model?

The published proofs are bounded experiments and specialized capabilities. They do not establish general intelligence or broad frontier-model equivalence. Choose a specific task and compare it directly.

Can Yodi use durable private memory?

A documented DynaCapDB integration checks authorized retrieval after restart. The public memory evaluation guide explains the small fixture and what a larger deployment still needs to verify.

Does local AI automatically keep every part of a workflow private?

No. Inspect the exact model route, provider calls, storage, logs, and application integration. Local execution is one property; it does not establish all of an organization's data-handling requirements.

From an idea to your first evaluation

Build on what you understand.

Start with Yodi's public evidence. If it fits your team's problem, compare source memberships or tell us what you would like to evaluate.

These guides are free to read. Private source releases follow the membership license. Compare plans and team seats.

Continue exploring